Search papers, labs, and topics across Lattice.
This paper introduces a unified framework for enhancing diversity in open-ended generation tasks by characterizing test-time diversity methods based on the source of diversity introduced during generation. The authors propose a novel approach that generates diverse intermediate specifications, which are then used to condition final outputs, leading to improved diversity without sacrificing quality. Their findings reveal that both the diversity of the sources and the effectiveness of their transmission to the output are critical factors in achieving meaningful variation in generated responses across multiple models and tasks.
Injecting diversity at the specification level can significantly enhance output variety while preserving quality, challenging conventional methods that often fall short.
Open-ended generation tasks often require a set of meaningfully different outputs, yet large language models often produce similar generations. Existing test-time diversity methods operate at different stages of generation with varying effectiveness, but it remains unclear what design choices lead to meaningful diversity in the output. We introduce a framework that characterizes test-time diverse generation methods by the diversity source introduced during generation and provide a transmission score for measuring how effectively variation in the source reaches the final output. Guided by this framework, we propose fully automated specification-level generation methods that first generate diverse intermediate specifications and then condition on them to produce final responses. Across five open-ended tasks and four backbone models, specification-level injection improves output diversity over test-time baselines while maintaining comparable quality. Our analysis shows that successful diversity injection depends on both the diversity of the sources and their transmission to the output, highlighting source design and source-to-output realization as two key levers for building more diverse generation systems.
UW